Rubik’s Optical Neural Networks: Multi-task Learning with Physics-aware System and Algorithms
TimeTuesday, July 12th6pm - 7pm PDT
LocationLevel 2 Lobby
Event Type
Networking Reception
Work-in-Progress Poster
DescriptionDue to the challenge of implementing reconfigurability, deploying multi-task learning (MTL) algorithms on Diffractive Deep Neural Networks (D2NNs) requires re-building and duplicating the physical diffractive systems, which significantly degrades the energy and cost efficiency in practical application scenarios. This work presents a novel D2NNs architecture, namely RubikONNs, which utilizes the physical properties of optical systems to encode multiple feed-forward functions by physically rotating the hardware similarly to rotating a Rubik's Cube, with two novel training algorithms. Our experimental results demonstrate more than 4x improvements in energy and cost efficiency with marginal accuracy degradation compared to the state-of-the-art approaches.